Bias and Variance as Error Sources

Stable wrongness versus unstable sensitivity

In this context, bias is systematic error from restrictive assumptions. A straight line fitted to a strongly curved relationship has high bias. Variance is sensitivity to the particular training sample. A deep decision tree whose structure changes dramatically after a few rows change has high variance.

Increasing capacity often lowers bias and raises variance. More representative data, bagging, and stronger regularization can reduce variance; better features or a more flexible model can reduce bias. This bias-variance trade-off is a diagnostic lens, not a rule that every intervention moves exactly one dial.

high bias clues:     train poor, validation poor, errors systematic
high variance clues: train excellent, validation weaker, seed/fold scores unstable

Scenario: A linear model misses the same nonlinear threshold in every fold - likely bias. A deep tree alternates between excellent and weak validation scores depending on the fold - likely variance.

Tip: Inspect fold-to-fold spread, not only average cross-validation score. Instability is evidence that deployment performance may depend heavily on which examples happened to be collected.